.png){kind=icon}
.png){kind=icon}
Return
GKTLF: a geographical knowledge-guided transfer learning framework for cross-regional DEM super-resolution
.png){kind=icon}
PRE
AI
X
M
陈
C
DOI：10.1080/13658816.2026.2704882.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Existing methods for digital elevation model (DEM) super-resolution cannot stably generate high-resolution DEMs, especially when transferred to target (test) regions with significant terrain spatial heterogeneity compared to the source (training) region. To address this problem, we propose a geographical knowledge-guided transfer learning framework (GKTLF) for cross-regional DEM super-resolution. Specifically, a cross-regional transfer strategy was designed to preserve terrain features in the source region and extract terrain features from the target region. Additionally, a geographical feature guidance strategy was introduced to integrate geographical knowledge, including masked terrain skeleton lines and slope features. Finally, a terrain similarity measurement method was developed to quantify terrain spatial heterogeneity. The experimental results showed that the proposed GKTLF achieved superior performance over the baseline networks across four test regions in China. Compared to enhanced super-resolution generative adversarial networks (ESRGANs), the GKTLF-enhanced ESRGAN achieved improvements in accuracy of 35.8, 25.2, 19.7, and 4.6% with full-sample pre-training and 31.2, 26.5, 20.4, and 2.5% when only 40% of samples were pretrained across the four test regions. This study can help mitigate significant cross-regional terrain spatial heterogeneity and provide a valuable reference for related research.
Keywords:
Cross-regional digital elevation model super-resolution
geographical knowledge
terrain spatial heterogeneity
terrain similarity measurement
transfer learning
Journal
IF:
5.1
Papers:
2.7K
Citations:
9.3K
Organization
Cited Papers